There are three layers to understand: base models, instruction-tuned models, and user-fine-tuned models. Most fine-tuning projects start from the second layer — not the first.
Base models — Trained on raw text with next-token prediction. Understands language deeply but does not follow instructions or behave helpfully. Not useful directly for most applications. Examples: LLaMA base, Mistral base.
Instruction-tuned models — Base model fine-tuned by the provider on instruction-following data. Helpful, conversational, safe. What you use via API today. Examples: GPT-4, Claude, Gemini, LLaMA-3-Instruct.
User fine-tuned — Instruction-tuned model further adapted for a specific application. Retains helpfulness from layer 2, gains specialized behavior from your training data.
Fine-tuning from a raw base model means you also have to teach it to be helpful and follow instructions — significant extra work. Starting from an instruction-tuned model means you're only adding your specific behavioral adaptations on top of an already-capable foundation.
Download the weights and fine-tune directly. Full control over training and the fine-tuned model. No per-token inference cost after training. Requires GPU infrastructure. Examples: fine-tune with Hugging Face Transformers or Axolotl.
Training happens on the provider's infrastructure. Fine-tuned model accessed via API — you don't own the weights. Higher per-token cost but zero infrastructure overhead. Easier to start.
Ask the AI assistant about the difference between base and instruction-tuned models, or how to choose a starting point for fine-tuning.